Synopsis

Keywords: PET/MR, Brain, attenuation correction

We evaluated the accuracy of a deep learning-based PET/MR attenuation correction (AC) method with vendor-provided high-resolution Dixon in- and opp-phase images as inputs (DL-HiRes). We found that the DL-HiRes AC method significantly outperformed the vendor-provided skull model AC method for both 16-channel head-neck coil and 32-channel head coil (p<0.001). Moreover, the DL-HiRes method had similar AC accuracy using different head coils.

Introduction

Deep learning-based methods have demonstrated improved accuracy for PET attenuation correction (AC) compared to conventional methods(1-4). We recently developed a 3D patch-based residual UNet (ResUnet) structure to generate pseudo-CT (pCT) using various MRI scans for PET/MR AC(4). The DL-DIXON method provided excellent PET/MR AC accuracy using standard Dixon in- and opp-phase images as inputs(4).
MR scanner upgrades, including software and hardware upgrades, are often introduced by the vendor. For example, a high-resolution (HiRes) Dixon scan was employed by the vendor to generate skull model μ-maps under the latest scanner software version (VE11P). In this skull model, the pre-saved skull bone is inserted through landmark matching. The skull model PET/MR AC is substantially better than the previously vendor-provided Dixon PET/MR AC(5). Moreover, a new 32-channel head coil was recently introduced to improve MR image quality for the Siemens Biograph mMR. Compared to the original 16-channel head-neck coil, images acquired using the 32-channel head coil have a higher degree of spatial signal variations. It remains unknown whether the MR coil may impact the performance of PET/MR AC. In this study, we trained a deep neural network using HiRes Dixon images (DL-HiRes). We evaluated the performance of DL-HiRes and the vendor-provided skull model using both the 16-channel head-neck and 32-channel head coils.

Methods

Tri-modality PET/MR (on a Siemens Biograph mMR 3T scanner) and PET/CT (on a Siemens Biograph mCT or Siemens Biograph Vision scanner) images were acquired from two groups of participants with IRB approval and informed consent. Group A consists of 24 participants (median [Interquartile range] age: 71.5 [68.8, 74.2], 14 female) with MR data acquired using the 16-channel head-neck coil. Group B consists of 24 participants (38.5 [28.8, 46.5], 19 female) with MR data acquired using the 32-channel head coil.
Low-dose CT images were acquired at 120kVp from all participants. PET images were acquired using ¹⁸F-Florbetapir tracer for Group A, and ¹¹C-PiB (N=3) or ¹⁸F-FDG (N=21) tracer for Group B. In- and opp-phase Dixon images were acquired using vendor-provided high-resolution Dixon AC scan (TE1/TE2/TR=1.28/2.51/4.14ms, FA=10°, voxel size=1.3x1.3x2mm, acquisition time=39sec).
The previously published DL-DIXON network trained using vendor-provided standard Dixon in- and opp-phase images (TE1/TE2/TR=1.23/2.46/3.6ms, FA=10°, voxel size=2.6x2.6x3.1mm, acquisition time=19sec) was used as the initial model for this study(4). The DL-HiRes networks were trained using in- and opp-phase high-resolution Dixon images as inputs. A network for 16-channel coil data was retrained using Group A, and a network for 32-channel coil data was retrained using Group B. We used three-fold cross-validation to evaluate the neural network performance: each time, 12 participants were used for training, 4 participants were used for validation, and 8 participants were used for testing.
The trained networks were applied to the high-resolution Dixon images to generate pCT images. CT and pCT images were converted to attenuation maps through a piecewise linear scaling(6). PET images were reconstructed using e7Tools with three attenuation maps: (1) gold standard CT-based μ-map. (2) DL-HiRes μ-map and (3) vendor-provided skull model μ-map derived using the high-resolution Dixon scan(7).

Results

The pCT mean absolute error (MAE) of DL-HiRes using the 16-channel coil is not significantly different from that using the 32-channel coil. (71.41±8.58HU vs 70.16±8.30HU, p=0.74; Figure 1). Figure 2 shows the CT μ-map, the DL-HiRes μ-map, and the skull model μ-map using the 16-channel coil (Figure 2A) and the 32-channel coil (Figure 2B) in representative participants.
Two participants in Group A were excluded from PET evaluation due to excessive artifacts on the vendor-provided skull model μ-maps. Figure 3 illustrates the PET mean relative error (MRE) across the participants. As shown in Figure 4A, small PET MRE (within ±2.5%) was consistently observed using DL-HiRes. The PET mean relative absolute error (MRAE) of DL-HiRes was significantly smaller than that of skull model for both the 16-channel and 32-channel coils (p<0.001), while the same method performed similarly between the 16-channel and 32-channel coils (p>0.15) (Figure 4B). Using the 16-channel coil, 99.68% [99.35%, 99.84%] and 92.31% [90.59%, 94.76%] of brain voxels had PET MRAE less than 10% for DL-HiRes and skull model, respectively (Figure 5A,B). Using the 32-channel coil, 99.97% [99.89%, 100.00%] and 92.77% [89.54%, 95.94%] of brain voxels had PET MRAE less than 10% for DL-HiRes and skull model, respectively (Figure 5C,D).

Discussion

We demonstrated that the DL-HiRes method outperformed the skull model method using either the 16-channel or 32-channel coil. A recent consensus paper suggested using “voxelwise MRAE below 10% in at least 90% of the voxels in the brain mask” as a qualification criterion for PET/MR AC methods(5). The DL-HiRes method passed this criterion for all participants, while about 25% of participants failed this criterion in the skull model method using either the 16-channel or the 32-channel coils. Moreover, the DL-HiRes method had similar PET/MR AC accuracy using either the 16-channel or the 32-channel coil.

Conclusion

Our results suggested that the DL-HiRes AC method is accurate, and has consistent performance when using different coils, supporting its use in PET/MR clinical studies.

Acknowledgements

This study was supported by NIH 1R01NS082561, 1P30NS098577, 5R01CA212148, P50AG05681, P01AG026276, P01AG003991, UL1TR000448, 1P30NS098577, T R01NS103988 and Siemens Healthineers. Avid Radiopharmaceuticals (a wholly owned subsidiary of Eli Lilly) provided the doses of 18F-Florbetapir and partially funded the cost of the PET scans.